KR101609644B1 - weighted least squares-based intrafield de-interlacing Method - Google Patents

Abstract

A minimum square-based intra-field deinterlacing method according to an exemplary embodiment of the present invention includes a step of estimating a statistical model of an original image for a conditional probability model with a Gaussian distribution, Setting a local window and deinterlacing the local window based on the statistical model analysis.

Description

[0001] The present invention relates to a weighted least squares-based intrafield de-interlacing method,

The present invention relates to a deinterlacing technique, and more particularly to a weighted least squares-based intrafield de-interlacing algorithm with a given weight.

Interlaced scanning has been used for video signals to reduce bandwidth and is widely used as the standard for TV broadcast systems in the form of NTSC, PAL, SECAM, and some HDTV formats. However, interlaced scanning produces some artifacts such as line flicker and sawtooth edges. To solve these problems, several deinterlacing algorithms have been proposed in the field to reduce the above drawbacks. Deinterlacing is a technique for generating an interlaced low-resolution image as a full-size high-resolution image.

De-interlacing is a simple approach to general upsampling problems. Since it can be performed by double vertical upsampling. However, such upsampling is only suitable when the video signal satisfies the Nyquist criterion. There were a number of de-interlacing schemes that provided a good balance between computation and hardware costs and the quality of the resulting images.

In general, de-interlacing is classified into two types. One is intrafield (spatial) de-interlacing, which accepts only a single field. The other is an interfield (spatio-temporal) de-interlacing method that accepts multiple fields. The MC deinterlacing method tries to regenerate motion trajectories with high correlation. However, the MC based approach requires high accuracy motion information.

Spatial deinterlacing is the simplest method and does not require high hardware costs. These schemes use a high correlation between the pixels in the current field and the pixels to be filled. Therefore, spatial methods are widely used in real-time applications.

There are a number of spatial deinterlacing schemes with some complexity and performance. For example, the line averaging method is an approach which is not complicated as a method of estimating an empty pixel with an average value of the intensity of the upper and lower pixels. Line averaging is preferred for its simplicity. However, since the vertical resolution of the input image is reduced by half until the image is regenerated, this provides a cause to weaken the detail of the image being de-interlaced.

To alleviate this problem, an edge-based line averaging (ELA) scheme appears. The ELA approach used edge orientation associations to regenerate empty pixels between two neighboring lines in an interlaced image. The ELA exhibits acceptable acceptable performance with low complexity and eliminates blurring artifacts in line averaging where the edges are properly determined. However, ELA gives low visual results in areas of high spatial frequency and inaccurate edge directions.

EELA has improved edge direction measures to better determine directional spatial associations.

Several upgraded versions of EELA have been introduced, including Directional Oriented Interpolation (DOI), Low Complexity Interpolation Deinterlacing (LCID), Content Analysis Based Approach (Modified Edge Based Line Average (MELA)), New Edge Dependent Deinterlacing (NEDD) Map-based deinterlacing (EMD), covariance-based accommodative deinterlacing (CAD), and edge-based association acceptance (ECA). For both intrafield deinterlacing using bi-directional filtering interpolation and deinterlacing, an adaptive bidirectional filter has been proposed using local intensity histogram screening and a generalized fuzzy computation approach.

The DOI accepts upper and lower spatial direction vectors to better estimate the direction of the highest spatial associativity. Using these vectors, the DOI reduced the chance of false decisions for the most relevant spatial directions. The LCID scheme represents an additional measure of effectively estimating spatial relationships for the main purpose of mitigating decision errors for subsequent interpolation. The MELA includes another vertical measurement and considers spatial relationships in three directions including vertical and two rectilinear directions. The NEDD uses both the difference and the edge pattern to detect the exact edge direction. The EMD introduces three edge direction vectors of 45 degrees, 90 degrees, and 135 degrees calculated by the edge map to obtain better resolution in the edge direction. According to the edge direction of the direction vector, weights are calculated for each edge direction. These weight values are multiplied by the candidate deinterlaced pixels to make the estimation of the reproduced scene successful. The CAD scheme accommodates adjacent pixels aligned in the direction of the edge region to obtain the optimal coefficients. Then, based on the directional geographical duality, an optimal interpolation coefficient for an adjacent pixel in a corresponding direction is estimated. The ECA is based on various edge orientations to find edges and is improved using the weighted sum of the ELA elements to facilitate interpolation results.

In the above-mentioned intrafield deinterlacing schemes, some methods such as line average, ELA, EELA, etc., are easy to implement due to their simplicity. However, these are difficult to achieve good visual quality due to low-level estimates using limited candidate edge orientation. CAD is an acceptable approach based on covariance estimation from low-resolution interlaced images to high-resolution deinterlaced images. CAD only degrades the image quality if the estimation is wrong.

From the existing deinterlacing algorithm, it can be concluded that the estimation quality of the deinterlacing algorithm is fundamentally dependent on its flexibility for various pixel structures over one image.

Conventional deinterlacing algorithms are efficient and easy to implement, but are vulnerable to restoring image quality and vulnerable to estimating the underlying structure of the original image, resulting in low visual quality and low PSNR.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to provide a deinterlacing algorithm that can implement a high quality in deinterlacing an interlaced image, estimate a fundamental image, increase visual quality and PSNR have.

According to an aspect of the present invention, there is provided a method of minimizing intra-field deinterlacing of a minimum square based on a weighted square, the method comprising: estimating a statistical model of an original image for a conditional probability model by a Gaussian distribution; Setting a weighted local window of a predefined predetermined size in the local window and deinterlacing the local window based on the statistical model analysis.

According to the weighted minimum square-based inner field deinterlacing algorithm according to the embodiment of the present invention, a high-resolution image is estimated from the interlaced low-resolution image by using the weighted minimum square error criterion of the interlaced image, It is possible to provide high deinterlacing performance in terms of visible quality.

1 shows an example of an image statistical model;
Figure 2 shows neighboring HR pixels of an HR pixel according to an embodiment of the present invention;
Figure 3 shows neighboring HR pixels of a LR pixel and neighboring HR pixels of an HR pixel according to an embodiment of the present invention;
4 shows neighboring LR pixels of an LR pixel;
FIG. 5 is a graph showing a matching ratio between an empirical histogram and a statistical model according to an embodiment of the present invention. FIG.
6 illustrates pixel locations within a local window in accordance with an embodiment of the present invention.
Figure 7 is a snapshot of a test image for performance evaluation in accordance with an embodiment of the present invention;
8 is a view showing a performance evaluation method;
9 is a view showing an original image and respective deinterlaced images thereof;
10 is a view showing an original image and respective deinterlaced images thereof;

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a weighted minimum square based inner field deinterlacing algorithm according to the present invention will be described in detail with reference to the preferred embodiments of the present invention with reference to the accompanying drawings.

In the present invention, a weighted minimum square (WLS) error criterion scheme is proposed to recover a fundamental original image. That is, the present invention proposes a deinterlacing algorithm using a minimum square (WLS) error criterion scheme with weighted interlaced images, which belongs to the class of a maximum a posteriori (MAP) framework.

The present invention uses statistical interpolation to make MAP issues neat, taking into account the influence and relationship between low resolution images and high resolution images. The present invention accepts a Gaussian distribution to approach empirical graphs to obtain a proper estimate of the statistical model. Utilizing the analysis of the statistical model, the present invention creates an estimation procedure for a weighted minimum square problem. The weights should be accepted by examining some statistical judgments and other characteristics of the image structure.

1. The proposed algorithm

A. MAP Estimator

First, the present invention introduces some notations necessary for describing the proposed algorithm. Let x be a given interlaced low-resolution (LR) image with the size of width * height / 2 and y be the high-resolution (HR) image to be interpolated with the size of width * height. x _m ∈ x and y _m ∈ y are the pixels of the LR and HR images. For the de-interlacing algorithm, the LR image is an odd column of the original image. Then, since x _m ∈ x implies x _m ∈ y, the HR image y is related to x. Therefore, the present invention can write y _m as x _m when y _m is in the LR image. The present invention can generate a deinterlacing problem with MAP estimation.

y _opt = arg max p (y | x)

               y Equation (1)

y _opt is the optimal estimate of the original image. Use Bayes rules to get:

y _opt = arg max p (y | x) p (y)

               y (2)

From equation (2), in order to estimate the optimum reproducibility, the present invention must obtain a statistical model of the original image and a conditional probability model.

For the natural image shown in Fig. 1 (a), it can be clearly seen that it is difficult to directly find the statistical properties of the image with the graph shown in Fig. 1 (b). It is well known that the residual image shown in FIG. 1 (c) has Gaussian statistical properties, which can be expressed as a graph of the residual image of FIG. 1 (d). The residual image is the difference between the original image and the smoothed image due to the average of the original image in the local window. The present invention can then estimate the previous statistical model as a Gaussian distribution. As follows,

식(3)

Equation (3)

When α _A in the range factor, σ _A is a weighting factor, A of factors are _{A = [a 1, a 2} , ..., a 8] of model parameters deulyigo, _N y is the y, as shown in FIG. 2 HR Neighboring pixels.

B. Proposed Deinterlacing Algorithm Using Error Weighted Weighted Minimum Squares (WLS)

For the conditional probability model, the present invention can make use of the above analysis and impart a characteristic of the conditional probability model as a Gaussian distribution.

식 (4)

Equation (4)

B is the range factor, σ _B is the weight factor, and y _x is the HR of the LR image x, where _B is the model factor, where B = [b ₁ , b ₂ , ..., b ₆ ] deulyigo neighboring pixels, x _y LR are the neighboring pixels on the HR image y in Figure (3).

The present invention continues to develop a MAP evaluator, and Equation (2) is equivalent to:

y _opt = arg max p (y | x) p (y)

= arg max log [p (x | y) p (y)]

_{= Arg max log [σ A exp} (-Ⅱy-Ay N 2 2/2 σ 2 A) σ B exp (-ⅡX-By X Ⅱ 2 2 + Ⅱy-Bx y 2 2 / 2σ 2 B)]

= arg max [log σ _A + (-IIy-Ay _N Ⅱ ² ₂ / 2σ ² _A ) + log _B + (-IIX-By _X Ⅱ ² ₂ + IIy-Bx _y Ⅱ ² ₂ / 2σ ² _B )]

= argmax [(- IIy-Ay _N II ² ₂ / 2σ ² _A ) + (- IIX-By _X II ² ₂ + y-Bx _y II ² ₂ / 2σ ² _B )]

= arg min [IIy-Ay _N Ⅱ ² ₂ / 2σ ² _A + IIX-By _X II ² ₂ + IIy-Bx _y II / σ ² _B ]

log _A and log _B are constants and they do not affect the maximum of p (x | y) p (y). The object of the present invention is to find the optimal y value by selecting the maximum value of p (x | y) p (y). Even if we subtract log _A and log _B from equation (5), there is no difference. Therefore, the present invention subtracts log _A and log _B to simplify the equation.

The main issue for successfully obtaining the best estimates is how to calculate the model factors A and B. A is obtained by a least squares approach assuming that the directional associations between LR pixels are similar to the directional associations between HR pixels.

식 (6)

Equation (6)

When X _S is a pixel vector containing all LR pixels in the local window S centered at x _m . X _N is a data matrix in which the k-th column is composed of eight LR neighboring pixels as shown in FIG. 4 (a). And W _A = diag [σ _A ^-2 ]. The solution of equation (6)

식 (7)

Equation (7)

To calculate the model factor B, the present invention includes another minimum square equation.

=arg _B min W_BX_S-Bx_N ⁽⁶⁾Ⅱ² ₂ 식 (8)

= arg_Bmin W_BX_S-Bx_N ⁽⁶⁾Ⅱ² ₂ Equation (8)

X _N is a data matrix in which the k-th column is composed of six LR neighboring pixels as shown in FIG. 4 (b). And WB = diag [σ _B ^-2 ]. The solution of equation (8)

=(X_N ^(6)TW_Bx_N ⁽⁶⁾)^-1(X_N ^(6)TW_Bx_S) 식 (9)

= (X_N ^{(6) T}W_Bx_N ⁽⁶⁾)^-One(X_N ^{(6) T}W_Bx_S) Equation (9)

It should be noted that the estimation of B assumes the following. That is, the interlaced pixel x _m has the same spatial direction and the same range, such that y _m is related to B in equation (5).

C. Selection of Weighting Factors

While the weight factor σ _A controls the width of the distribution, the range factor α _A controls the height of the distribution. For different range factors? _A and weight factors? _A , the matching has various characteristics, as shown in FIG.

In this aspect, the present invention describes an empirical graph of residual images of several natural images and compares the empirical graph with previous statistical models with various factors. It can be seen in Figure 5 that the width plays a more important role than the height. Because when the residual image is close to zero, the pixels are in the smooth area. In image processing, all interest is given to complex areas such as edges or textures at the end of the distribution. The end is controlled by the width. Moreover, it is noted that the range factor is inversely proportional to the weight factor. Equation (5) confirms that the fact that the range factor does not affect the MAP estimate and that only the weight factor obtains the best estimate is to be considered. Based on this analysis, it is assumed that the present invention only discusses weighting factors. To achieve the best match in the empirical graph, the present invention should use flexible factors with different values. In other words, for large residual values, the present invention may use large weighting factors, and for small residual values, the present invention may use small weighting factors. In an actual calculation, the present invention utilizes a weighting matrix W _A whose diagonal elements W _{A (i, i)} = σ _A ^-2 (m, i). Then, the weighting matrix W _{A (i, i)} will be inversely proportional to the weighting factor _A ² σ (m, i), and the weighting factor σ _A ² (m, i) must be proportional to the residual value. Therefore, W _{A (i, i)} should be inversely proportional to the residual value. Then, the present invention can define W _{A (i, i)} as follows using both filters.

식 (10)

Equation (10)

v _xm and v _xi are the pixel positions of x _m and x _i , and α and β are the range arguments.

For the weighting factor σ _B of the model factor B in Eq. (5), two elements || x By _x || ₂ ² and || y Bx _y || _{There are 2} ² . Therefore, the present invention should design two different weighting matrices W _B ¹ and W _B ² for the model factor B.

식(11)

Equation (11)

λ is a regularizing element that adjusts the contributions of model factors A and B, where α and β are equal to W _{A (i, i)} , and v _ym and v _yi are pixel locations of y _m and y _i . The HR pixels y _m and y _i are not known. The present invention then uses a line average to estimate y _m and y _i in equation (11).

The present invention defines the ratio of model factors A and B to.

식 (12)

Equation (12)

D. Implementation of the proposed algorithm

In the weighted minimum square-based de-interlacing method proposed above, the present invention estimates an HR pixel y ₁₃ utilizing a local window as shown in FIG. The reason for choosing a larger window size is to get better PSNR and more stable statistical properties. In the proposed method, the present invention utilizes local statistical properties to approximate the relationship between high resolution and low resolution pixels. The large window size assumes that more reliable statistical properties are involved. Based on the above discussion, the present invention can rewrite equation (5) as follows.

식 (13)

Equation (13)

When W _A is a 1 * 1 matrix, W _B ¹ is a 25 * 25 matrix, and W _B ² is a 6 * 6 matrix.

The present invention can also reduce equation (13) as follows.

식 (14)

Equation (14)

_{y L = [y 1, y} 2, ..., y 25] T and _{x L = [x 1, x} 2, ..., x 42] T is not known, local window, as shown in FIG. 6 W is a 32 * 32 diagonal matrix and its elements are composed of W _A , W _B ¹ , and W _B ² elements.

A matrix C having a size of 32 * 32 and a matrix D having a size of 32 * 42 are composed of factors A and B. According to equation (13), C and D can be expressed as C = [C ₁ C ₂ C ₃ ] ^T and D = [D ₁ D ₂ D ₃ ] ^T. In equation (13), C ₁ and D ₁ are related to the first term, C ₂ and D ₂ are related to the second term, and C ₃ and D ₃ are related to the third term.

C ₁ is the 6 * 25 matrix consisting of the factor A corresponding to the 6 connected neighbors HRPs in the k th column of LRP x _k in FIG. D ₁ is a 6 * 42 matrix consisting of a 6 * 6 matrix of 1s and a 6 * 22 matrix of 0s.

C ₂ is a 25 * 25 matrix composed of 1, D ₂ is a 25 * 42 matrix composed of factor A corresponding to the six connected neighbors LRPs of HRP y _{k in} the k th column in (3). C ₃ is a 1 * 25 matrix in which the k-th column is a factor B corresponding to six connected neighbors HRPs of HRP y _k , and D ₃ is a zero matrix of size 1 * 42. Using the LR interlaced image x, the present invention calculates all of the above factors and estimates the deinterlaced HR image y.

2. Simulation Results

Experiments performed under various settings and conditions are discussed below. Several numerical and visual results are presented for the proposed WLS-based deinterlacing (WLSD) scheme and the underlying schemes from recent documents. We demonstrate the superior performance of WLSD compared to existing methods, line averaging [8], ELA [9], EELA [10], NEDD [14], CAD [16], ECA [17] and FDD [18] The results of the simulation are based on eight representative test images, Chair (512 * 512) ... Caravel (768 * 512), test scenes, Bluesky (1920 * 1080) ... Akiyo (176 * 144) . These images are shown in FIG.

One objective quality metric (PSNR) was used to estimate the performance of the WLSD scheme. The deinterlacing rate was measured by measuring the time required by each scheme for deinterlacing the images. Structural Similarity (SSIM) is an attempt to measure subjective performance, using a perceptual model. Smaller SSIM values tend to be more error-prone in the estimated image, and therefore, visual quality is perceived to be much lower. Experiments were performed on an Intel Core 2 Duo CPU E8500 3.14 GHz environment using a MATLAB implementation.

A. Objective Performance Analysis

The present invention uses the PSNR metric for evaluating objective quality and has defined as follows.

식 (15)

식 (16)

Equation (15)

Equation (16)

img _org and img _src represent the original and regenerated images. All test images are converted to a vertically interlaced size, based on the system as in FIG. 8, at original size. And the regenerated images are compared with the original image.

Table 1 shows the PSNR and average error rate of the eight test images. It can be seen from Table 1 that the WLSD algorithm represents the optimal PSNR. The improvement in average PSNR is 2.3 / 2.75 / 2.28 / 2.10 / 3.35 / 2.13 / 1.86 dB compared to the existing algorithms, line average / ELA / EELA / NEDD / ECA / FDD / CAD.

Method LA ELA EELA NEDD ECA FDD CAD WLSD AWLSD Bench 30.63 31.52 31.92 32.00 30.33 30.96 31.69 34.76 34.18 Clock 35.58 34.91 35.83 35.71 34.11 35.91 35.98 38.32 37.94 Carrousel 33.15 33.36 33.81 33.88 32.19 33.76 33.85 35.49 35.16 Flower 37.34 37.03 37.46 37.35 35.82 37.67 37.57 40.42 39.99 Guitar 32.59 31.99 32.18 33.13 31.56 32.55 33.95 35.49 35.08 Butterfly 35.10 34.46 34.92 35.07 33.52 35.51 35.61 37.30 36.99 Window 34.37 34.18 34.56 34.57 33.66 34.60 34.87 35.98 35.75 Caravel 32.29 31.23 31.52 31.24 31.55 32.22 32.81 33.77 33.56 Bluesky 37.88 37.24 37.36 37.89 37.59 37.86 37.85 40.10 39.79 Raven 42.13 40.03 40.62 41.90 40.66 41.68 41.89 43.08 42.95 Forman 32.24 33.49 33.63 33.13 32.07 32.90 32.93 34.07 33.81 Akiyo 35.44 33.99 35.32 35.28 33.15 35.26 35.31 37.65 37.34 Average 34.90 34.45 34.92 35.10 33.85 35.07 35.34 37.20 36.88

PSNR metrics of deinterlaced images and video

The present invention now compares the performance of WLSD deinterlacing algorithms for other approaches in terms of computation time. Table 2 shows the time it takes to de-interlace one of the test images. The WLSD scheme is the slowest compared to other schemes. As can be seen in Table (1) and Table (2), the WLSD scheme provides the highest PSNR than other schemes, but requires more computation time. With advances in technology, CPU computation time is not currently a big issue. And there are many techniques to improve the CPU computation time.

Method LA ELA EELA NEDD ECA FDD CAD WLSD AWLSD Bench 0.015 0.218 0.265 5.351 0.580 2.491 13.908 41.901 4.772 Clock 0.021 0.202 0.234 5.351 0.565 2.195 13.642 41.655 4.422 Carrousel 0.021 0.187 0.249 5.397 0.565 2.163 13.892 41.832 4.489 Flower 0.015 0.202 0.234 5.396 0.565 2.242 13.705 41.679 4.538 Guitar 0.015 0.187 0.249 5.364 0.565 2.273 13.565 41.365 4.521 Butterfly 0.021 0.187 0.249 5.302 0.565 2.132 13.986 41.336 4.712 Window 0.220 0.343 0.349 8.201 0.831 4.356 15.336 41.399 4.388 caravel 0.220 0.359 0.361 8.081 0.819 4.568 15.056 41.338 4.376 Bluesky 0.062 1.654 1.825 10.316 4.212 17.016 84.661 186.721 19.232 Raven 0.026 0.733 0.585 5.318 1.965 6.562 38.017 112.268 11.500 Forman 0.001 0.078 0.095 0.577 0.215 0.828 8.916 18.608 2.077 Akiyo 0.001 0.033 0.039 0.268 0.107 0.386 1.086 2.596 0.517 Average 0.053 0.365 0.417 5.410 0.963 3.934 20.481 54.392 5.796

B. Subjective Performance Analysis

Subjective quality can be assessed for textures, edges, and other kinds of geographic details (diagonal, corner, quality patterns, etc.). To compare the subjective performance of the deinterlaced image, the present invention sees part of the perceived image quality of the Butterfly and Caravel in FIGS. 9 and 10. FIG. From the shapes shown, it can clearly be seen that the WLSD algorithm according to the present embodiment of the present invention provides superior subjective quality in the texture region compared to existing algorithms. The LA method provides low subjective performance when the edge direction is not perpendicular. The performance of the ELA is good when it is an estimate of the exact edge direction. However, there are some obvious artifacts due to incorrect orientation estimates. The EELA method measures only three edge directions, and incorrectly detected edge directions cause image quality degradation. NEDD shows poor visual quality in wrong direction estimates. Figures 9 (g) and 10 (g) show that ECA has comparable results with MELA and LCID. FDD represents more edge direction than EELA. However, due to the limitation of candidate directions, there are some obvious artifacts in the texture area.

Method LA ELA EELA NEDD ECA FDD CAD WLSD AWLSD Bench 0.9349 0.9460 0.9508 0.9499 0.9376 0.9459 0.9435 0.9600 0.9487 Clock 0.9820 0.9763 0.9807 0.9824 0.9738 0.9833 0.9794 0.9862 0.9843 Carrousel 0.9729 0.9658 0.9711 0.9730 0.9600 0.9747 0.9725 0.9776 0.9755 Flower 0.9716 0.9711 0.9759 0.9786 0.9749 0.9807 0.9764 0.9805 0.9765 Guitar 0.9696 0.9628 0.9692 0.9764 0.9610 0.9696 0.9720 0.9786 0.9745 Butterfly 0.9817 0.9771 0.9793 0.9820 0.9741 0.9824 0.9808 0.9847 0.9833 Window 0.9653 0.9618 0.9652 0.9658 0.9576 0.9664 0.9662 0.9723 0.9691 caravel 0.9398 0.9294 0.9359 0.9351 0.9254 0.9397 0.9398 0.9453 0.9428 Bluesky 0.9807 0.9780 0.9787 0.9808 0.9798 0.9808 0.9806 0.9849 0.9830 Raven 0.9845 0.9752 0.9782 0.9828 0.9748 0.9846 0.9812 0.9849 0.9847 Forman 0.9389 0.9458 0.9480 0.9456 0.9363 0.9434 0.9451 0.9504 0.9452 Akiyo 0.9759 0.9651 0.9735 0.9747 0.9541 0.9779 0.9770 0.9825 0.9795 Average 0.9665 0.9629 0.9672 0.9689 0.9591 0.9691 0.9679 0.9740 0.9706

CAD has the best performance in the conventional way. However, it still has artifacts. In contrast, the WLSD approach can be tailored to the image to produce better visual quality than other methods. Figures 9 (k) and 10 (k) show the visible performance of an adjustable WLSD (AWLSD). All results show that the weighted minimum square based deinterlacing scheme according to the embodiment of the present invention is superior to other intrafield deinterlacing schemes in terms of low complexity, objective and subjective quality.

In order to continuously assess subjective image quality, the present invention accepts a measurement, called SSIM, which is one of the well known methods of subjective image quality assessment. The SSIM index compares local patterns of generalized pixel intensities with luminance and contrast, and the present invention can measure the similarity of an original image and a regenerated image through SSIM. The focus of the SSIM index is to capture the loss of structure in images. Since the human visual system is highly adaptive, structural information can be extracted from the visible scene. Thus, measurement of structural similarity provides a good estimate of perceived image quality. The SSIM index can be expressed as:

식 (17)

Equation (17)

식 (18)

Equation (18)

M is the number of local image areas. The present invention finds from Equation (17) that if μ _o is equal to μ _R and σ _o is equal to σ _R , SSIM has the largest value of 1. Therefore, SSIM has a range from 0 to 1. When the SSIM approaches 1, the regenerated image becomes more similar to the original image. And vice versa. MSSIM is the average of SSIM. Thus, MSSIM has attributes such as SSIM. In other words, the larger the MSSIM value, the better the subjective image quality. The MSSIM is compared to the schemes in table (3). Table 3 also shows that the WLSD scheme has better MSSIM compared to existing algorithms.

C. Adjustable WLSD

From Table 1 and Table 3 and Figures 9 and 10, the present invention finds that the proposed algorithm shows promising objective and visible results when compared to existing algorithms. However, the present invention also observes, in table 2, that the proposed algorithm has a relatively high computational complexity due to the computation of the inverse of the large window size to obtain good PSNR and visual quality. To alleviate the computational complexity problem, the present invention provides a coordinated method of reducing the complexity of the proposed WLSD. Conventional methods such as LA, ELA, and EELA appear to have some artifacts in the edge area. However, these methods produce good results in smooth areas. Based on these observations, the present invention presents a coordinated WLSD (AWLSD). A coordinated WLSD (AWLSD) consists of two methods, WLSD and LA. The proposed WLSD method is applied only to the edge region occupying only a small portion in the natural image, and LA is used in the smooth region. To extract this edge region, the present invention uses a Sobel operation. Tables (1) and (3) show that AWLSD outperforms existing intrafield deinterlacing methods in terms of PSNR metrics. Moreover, it is noteworthy that in Table 2, the AWLSD reduces the CPU computation time by 89.3% compared to the WLSD. From the results of the present invention, it can be concluded from the above results that AWLSD provides excellent performance and optimal quality-rate tradeoff among the methods studied so far.

3. Conclusion

The present invention has shown a detailed analysis of a deinterlacing algorithm using a weighted minimum square (WLS). The present invention estimates a high resolution image from an interlaced low resolution image by using a weighted minimum square error criterion based on the MAP framework and statistical analysis. It shows that it provides the best deinterlacing performance in terms of PSNR and SSIM. The proposed algorithm has high computational complexity due to large window size and matrix operation. The most significant contribution of this approach is that the present invention not only achieves higher PSNR but also better visual quality. The WLSD CPU time is 2.655 times the CAD and the PSNR result is 1.54dB higher. WLSD and AWLSD generate satisfactory visual performance. In future work, the present invention will invest in how to reduce window size, how to maintain good properties, and how to utilize other effective ways to replace matrix operations.

The embodiments of the minimum square-based inner field de-interlacing algorithm of the present invention described above are merely illustrative, and those skilled in the art will appreciate that various modifications and equivalent implementations You can see that examples are possible. Therefore, it is to be understood that the present invention is not limited to the above-described embodiments. Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims. It is also to be understood that the invention includes all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

Claims (3)

In the de-interlacing method,
Estimating a statistical model of the original image as a Gaussian distribution for the conditional probability model;
Setting a weighted local window of a predefined predetermined size in the interlaced image; And
And deinterlacing the local window based on the statistical model analysis. ≪ Desc / Clms Page number 19 > The method according to claim 1,
Wherein estimating the statistical model utilizes local statistical properties to approximate the relationship between high resolution and low resolution pixels of the image. delete

Images